A Novel Visual Based Approach To News Video Parsing

With the rapid development of the Internet age, the explosion of amount of multimedia information and the rise of related technologies, news video parsing is one of the most worthy topic to enterprise or institute, which offers video services. The correct news stories analysis can be more effective to retrieval what we want. This paper will propose a news analysis system, which is based on visual characteristics of anchorperson scene, aims to provide personalized news service in streaming media platform.The first step is structured division of news video about news content parsing. Based on the assumption that each news story launches with an anchorperson segment and follows relevant news footage, the system conducts anchorperson detection which splits the news video into several successive segments.In order to detection anchor person effectively, the author integrates a number of key techniques and methods into system to achieve correct detection results. First of all, use shot boundary detection to segment news video into many shot sets, and then extract key frames to represent the shot information. Secondly, detect face area by face detection algorithm and extract image features including color feature, texture feature and local feature and so on. Merge those features mentioned above and mapping to feature descriptor space. Make use of distance metric to measure dissimilarity between two face images. Next, employ an unsupervised and graph-theoretic-based clustering way, including MST and AHC, to get similar images together. The four clustering results selection criteria discriminate the anchor category and the non-anchor category. Hereafter, label timestamp of anchor scenes and segment news video into structured content. The fourth criterion is proposed by temporal distribution of anchor person and can handle appearance of multi-anchors.After structured segmentation, specific to each content-independent story unit, this paper presents a method of tagging the dominant characters in story unit. The aim of this investigation is to offer characters and content retrieval and indexing.The effectiveness and robustness of the proposed system are demonstrated by the evaluation on approximately21hours of news programs from France Channels and China Central Television Channels. Besides, the system achieves F-measure96%on average.